Agentforce
Koa: Salesforce built its own CRM model, and the pitch is fewer wrong actions
Salesforce announced Koa at Dreamforce, its first CRM reasoning model, post-trained on NVIDIA Nemotron 3 Super. The headline number is not a benchmark leaderboard, it is three times fewer errors on CRM actions. Here is what that claim means, where a purpose-built model beats a bigger general one, and how to slot Koa into the model menu you already have.
Most Agentforce agents do not fail because the model wrote a clumsy sentence. They fail because the model called the wrong action, updated the wrong field, or routed a case to the wrong queue. Those wrong-action failures are what Salesforce built Koa to cut.
Koa is Salesforce’s first CRM reasoning model, announced at Dreamforce on September 15. It’s not a general chat model with a Salesforce logo on it. It was built by post-training NVIDIA’s Nemotron 3 Super on a synthetic dataset modelled on how CRM work gets done. The claim that matters for anyone who runs agents in production is narrow and specific: on CRM actions, Koa makes about three times fewer errors than the leading general models.
My read: for the boring, high-volume actions that make up most of a service or sales agent’s day, a model trained on those actions is worth trying the moment it reaches your region. Keep a frontier general model for the open-ended reasoning. The decision is not “switch everything to Koa,” it is “which agent, and which subagent, should run on it.”
What Koa is
A reasoning model is the part of an agent that decides what to do next: read the request, plan the steps, pick the tool, fill the arguments, check the result, decide whether it is done. In Agentforce that job belongs to the Atlas reasoning engine, and the LLM sits inside Atlas as the thing that does the classifying and planning.
Until now that LLM was always someone else’s. You could pick a model from a menu, but every option was a general-purpose model that happened to be good at reasoning. Koa is the first one Salesforce trained itself, for this job.
The base is Nemotron 3 Super, one of NVIDIA’s open reasoning models. Salesforce post-trained it on a proprietary synthetic dataset built to reflect the reasoning, tool use, and decisions that CRM work involves: generating a lead, qualifying an opportunity, resolving a case, scheduling the follow-up. No customer data went into that training. The corpus is synthetic, modelled on patterns from Salesforce’s own decades of CRM deployments rather than on any org’s records.
The design choice is the interesting part. Koa is not trying to know more facts than GPT or Claude. It’s trying to be more reliable at a small, well-defined set of moves.
The error-rate claim, and how to read it
Salesforce measures Koa on its own benchmark, CRM Bench. The tasks are the ones an agent does all day: update an opportunity, route a case, schedule a follow-up. On that benchmark, Salesforce says Koa matches or exceeds leading models on CRM actions with three times fewer errors.
Take the framing seriously and the number with the usual caution. A vendor’s own benchmark is a vendor’s own benchmark, and “leading models” is not named. What CRM Bench measures is not eloquence. The metric is whether the model picks the right action and fills its parameters correctly, which is exactly where general models trip in an Agentforce agent.
I have watched a capable general model classify a topic perfectly and then pass a malformed record ID to the update action, or invent an optional parameter that the API rejects. Those are not hallucinations in the sentence sense. They are action errors, and they are the ones that make an agent look untrustworthy to the business, because they change data.
So the way to evaluate Koa is not to trust the three-times figure. Run your own worst actions against it. Take the ten actions your current agent gets wrong most often, build them into a test set with the Testing Center or your own test rig, and compare Koa against whatever you run today on that exact set. The claim is testable in your org. Test it there.
Where a purpose-built model wins, and where it does not
A smaller model trained on a narrow task tends to beat a bigger general model inside that task and lose everything outside it. That trade-off should shape where you put Koa.
Koa is the right bet for an agent whose job is mostly deterministic CRM operations. A service agent that reads a case, checks entitlement, updates status, and posts an update. A sales agent that logs activity, updates the opportunity stage, and books the next meeting. High volume, repetitive, action-heavy, and punished hard by a wrong tool call.
A general frontier model is still the better bet where the work is open-ended: summarising a messy multi-thread conversation, drafting a nuanced customer reply, reasoning across documents where the answer is not a CRM action at all. That’s also where Claude running through Claudeforce earns its place, and where you would not expect a CRM-tuned model to lead.
The per-subagent model override matters more now than it did. As I covered in the model selection guide, you can already run different subagents on different models. Koa makes that a real architectural lever rather than a nice-to-have. Route the deterministic action subagent to Koa and the free-text reasoning subagent to a frontier model, in the same agent.
A model trained on your actions beats a bigger model that has read more of the internet, but only inside those actions. Outside them it knows less.
The trust boundary is the quiet headline
The part your security team will care about is not the benchmark. It’s where Koa runs.
Salesforce controls Koa’s weights and runs the model entirely inside its own infrastructure. No customer data crosses the trust boundary during training, because training used synthetic data, and none crosses during inference either, because the model runs in Salesforce’s environment rather than being called out to a third-party API.
Compare that to the shape of running a general model. Even with the Einstein Trust Layer masking and zero-retention agreements in front of a third-party model, your prompt still leaves Salesforce to reach that model’s endpoint. The Trust Layer is real and it does its job, but it is a set of controls around an external call. Koa removes the external call.
For a regulated org that has spent months getting comfortable with a third-party model’s data handling, “the model runs inside Salesforce and never saw customer data” is a shorter conversation. That, as much as the error rate, is the reason Koa exists. It’s Salesforce’s answer to the data-residency and model-provenance questions that slow enterprise agent rollouts.
It does not remove your grounding obligations. Koa still answers from what you retrieve. A model that never hallucinates a sentence can still act confidently on bad grounding, so the Data 360 foundation underneath the agent matters exactly as much as it did before.
What you can do today, and what to wait for
Koa is available to select pilot customers now, with general availability expected in Winter 2026 in US regions first. That timing has two consequences worth planning around.
If you are not in the pilot, you cannot run Koa yet, and if you are outside the US you will wait longer. So the useful work now is preparation, not migration. Build the test set. Instrument your current agent so you know its real action-error rate today, because you cannot measure a three-times improvement against a number you never captured.
When Koa reaches GA, treat the switch like any other model change: a change that can break a working agent. Swapping the model underneath an agent changes how it classifies topics and plans, and an agent tuned against one model’s behaviour can regress against another. Move one subagent, regression-test it against your set, then widen. Do not flip the whole fleet on announcement day.
Watch the credit math too. Koa’s pricing was not detailed at launch, and model choice interacts with Flex Credits and the reasoning loop that drives most of your consumption. A cheaper or faster model that needs fewer retries because it gets the action right the first time can cost less per resolved case even if the per-token price is similar. Cost per resolved case is the number to compare, not cost per token.
The bigger pattern
Koa is the first CRM-specific model, and it will not be the last. The industry spent three years assuming the winning move was to call the biggest general model available. For action-heavy enterprise work, the evidence is starting to point the other way: a smaller model trained on the exact task, run close to the data, beats a general giant on reliability and on the compliance conversation at once.
If you build agents on Salesforce, that shifts your job. Model selection stops being “use the newest frontier model” and becomes “match the model to the subagent’s job.” Koa gives you a purpose-built option for the part of that job where reliability, not brilliance, is what the business is paying for.
Understanding the basics
What is Salesforce Koa?
Koa is Salesforce’s first CRM reasoning model, announced at Dreamforce 2026. It was built by post-training NVIDIA’s Nemotron 3 Super on a synthetic dataset modelled on CRM work, and it runs inside Salesforce’s own infrastructure to power reasoning and action selection in Agentforce agents.
Is Koa better than GPT or Claude?
Not in general. Salesforce’s claim is specific: on CRM actions, measured against its CRM Bench benchmark, Koa makes about three times fewer errors than leading general models. For open-ended reasoning and free-text drafting, a frontier general model is still likely to lead. The point of a per-subagent model override is that you do not have to choose one for everything.
Was customer data used to train Koa?
No. Salesforce says Koa was trained on a proprietary synthetic dataset rather than customer records, and that no customer data crosses the trust boundary during training or inference because the model runs inside Salesforce’s own infrastructure.
When is Koa available?
Koa is in pilot with select customers now, with general availability expected in Winter 2026, starting in US regions. Availability in other regions was not dated at launch.
If you build agents on Salesforce
Instrument your current agent’s action-error rate now, before Koa reaches you. The whole Koa pitch is a number you cannot verify without a baseline. If you want a second pair of eyes on which of your agents should move to a CRM-tuned model and which should stay on a frontier one, talk to us, or start with the Agentforce readiness assessment.
